What ZUMA-7, TRANSFORM and BELINDA measure about the temporal architecture of cell therapies, rather than about CAR biology
A doctrinal exploration of interventions whose manufacturing competes in speed with the disease itself. It transposes to treatment the question raised for predictive models in «Applicability Domain Measures a Proximity, Not a Capacity», and intersects «How Do You Attribute Performance in a Complex Socio-Technical Pipeline?» on the handling of post-randomization variables. The full formalism, the decomposition of delay determinants and the selection scenarios are in the PDF alongside.
Liso-cel and tisa-cel share the 4-1BB costimulation domain. TRANSFORM reports a median EFS of 29.5 months against 2.4 (HR 0.375). BELINDA reports 3.0 against 3.0, HR 1.07, non-significant. The CAR-biology account is available and comfortable, but it requires the same costimulation domain to yield two opposite verdicts. The temporal-architecture account asks for less. In BELINDA, tisa-cel imposes 52 days of manufacturing on a population selected at apheresis as last-resort salvage, with permissive bridging in 83 to 97% of patients. By day 52 the treated patient is no longer the randomized patient: he is what the disease made of him while the product was being manufactured. What the trial measures at that point is not the inferiority of 4-1BB, it is an exit from the performance domain before administration. TRANSFORM, with short manufacturing and favorable bridging selection, measures the converse on neighboring biology. ZUMA-7 validates axi-cel with 2-year EFS of 41% against 16% (HR 0.40) without having to settle the question. This reading does not prove CD28 and 4-1BB equivalent. It establishes something more modest and more useful: if biology differs, the system’s temporal architecture can compensate, and if the temporal architecture fails, no biology recovers it.
Three objects suffice to formalize the problem. The patient’s clinical state S_i(t) is a continuous trajectory, not a binary eligibility status. The delay τ_ij is a distribution F_ij characterized by its expectation, variance and P90, not by a single value. The response function R_j(S) assigns each treatment a performance that degrades progressively with the patient’s state, with no sharp boundary. From these follow expected utility V_ij = E[R_j(S_i(t0 + τ))] and the decision j* = arg max V_ij. The performance domain Ω_j(u*) = {S : R_j(S) ≥ u*} derived from it depends on the utility threshold chosen: it is a conventional region, not an absolute biological property of the product. Observed delay is not a sum but a composition, τ = g(τ_strategy, τ_patient, τ_system, τ_stochastic), with simultaneous interacting components: an infection occurs during manufacturing without adding linearly to it, and bridging may itself be caused by a production delay. The methodological consequence is direct. Under a treatment-policy estimand, delay is part of the assigned strategy’s effect whatever its determinant, and it is not neutralized statistically. For a causal analysis comparing delivery architectures, the decomposition becomes indispensable, since a fraction of the delay reflects the patient rather than the design. Naively adjusting a post-randomization variable without a target estimand and an explicit DAG corrects nothing, it relocates the bias.
Three operational consequences follow, and only one is painless. The first is measurement: current KPIs report vein-to-vein or apheresis-to-infusion, both blind to the decision-to-apheresis segment where the window is most often lost. The missing KPI is decision to effective administration, reported as a full distribution, because P90 predicts the share of temporal failures better than the median does. A treatment at median day 25 with P90 day 60 can be less useful than one at median day 30 with P90 day 35. The second is economic: in a cell manufacturing system run near saturation, the right tail of the distribution explodes, and the associated clinical loss grows faster than the utilization gain. Reserve capacity is not an inefficiency to eliminate, it is clinical capacity, and the optimal occupancy rate C* sits below 100% for reasons that are simultaneously clinical and financial. The third concerns selection: the ratio Θ = E[τ] / time of clinical evolution classifies the situations where timing becomes decisive, and this must be stated plainly, Θ is a dimensional heuristic and not a validated biomarker, its denominator remaining to be defined precisely. Used qualitatively, it is enough to decide. In ultra-fast disease, a bispecific available within days genuinely competes with an autologous CAR-T at 28 days, provided T cell fitness is acceptable, a parameter not routinely measured at the moment the decision is taken. In slow disease the temporal advantage vanishes and autologous CAR-T regains the edge on durability. Bispecifics therefore do not dominate, they compete through V_ij, which is intellectually more honest and clinically more usable than an announced succession. The validity domain of all this must be stated explicitly: the framework holds where Θ approaches or exceeds 1, meaning CAR-T, TIL, neoantigen vaccines, adaptive radiotherapy, transplantation, and it contributes nothing where delay is negligible against disease kinetics. One reformulation then remains, and it fits in a line. The relevant treatment is not the best treatment, it is the best reachable treatment, the one that reaches the patient while the patient is still inside its performance domain.
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.